Baum Welch Learning allows a HMM to be constructed from a series of sequence
observations. This implementation of Baum Welch scales and is not as
susceptible to underflows in long sequences of data as the regular Baum Welch
Baum Welch requires a starting point. You should create a HMM that has a
reasonable guess as to the observation and transition probabilities. If you
can make no such guess, you should consider using KMeans training.
L. E. Baum, T. Petrie, G. Soules, and N. Weiss,
"A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains"
, Ann. Math. Statist., vol. 41, no. 1, pp. 164-171, 1970.
Hidden Markov Models and the Baum-Welch Algorithm, IEEE Information Theory
Society Newsletter, Dec. 2003.